Intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis

Abstract

The present application relates to a kind of dual-energy X-ray image fusion and R value analysis intelligent mineral sorting method, including dual-energy X-ray image acquisition and pretreatment, multi-band image fusion, R value calculation and grade analysis, sorting model construction, and fusion decision, the quality of input image is improved by image preprocessing, the effectiveness and precision of R value estimation have been improved, multi-band fusion method effectively enhances the effect after ore dual-energy image fusion, based on the estimation method of R value analysis can realize the rapid estimation of ore grade, effectively expand the scale of image feature data, finally, R value estimation result and sorting model output fusion decision are combined, realize the efficient, accurate intelligent sorting of mineral.

Description

Technical Field

[0001] This invention relates to the field of mineral sorting technology, specifically to an intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis. Background Technology

[0002] Mineral resources are a vital foundation of the national economy, and their efficient sorting is crucial for improving resource utilization. However, traditional sorting methods suffer from low efficiency, high energy consumption, and significant pollution. X-ray transmission (XRT) sorting technology, as a non-destructive testing method based on internal composition characteristics, has received widespread attention in recent years. Dual-energy XRT, in particular, can effectively eliminate thickness interference through the ratio of high and low energy signals, theoretically enabling more accurate mineral identification and grade estimation. However, existing dual-energy XRT sorting technologies still have significant shortcomings: first, image fusion often uses fixed weights or general algorithms, failing to adapt to the energy spectrum characteristics of different minerals, resulting in poor sorting effects for low-contrast ores; second, R-value analysis is easily affected by noise and matrix effects, limiting its accuracy; and third, the system lacks deep integration with intelligent algorithms, making adaptive optimization difficult. These problems severely restrict the promotion and application of this technology in the sorting of complex non-ferrous metal ores. Therefore, an intelligent and efficient mineral sorting method has become a key research focus. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis. This method can adaptively perform image fusion processing for different mineral characteristics, improve the recognition accuracy of low-contrast ores, and solve the problem of insufficient training samples, thereby achieving efficient and accurate automatic mineral sorting.

[0004] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis, comprising the following steps: (1) High-energy and low-energy images of the ore were acquired simultaneously using a dual-energy X-ray transmission imaging system. The acquired images were preprocessed to extract images of the ore region. (2) After the extracted ore area image is processed by guided filtering, multi-band decomposition is performed to obtain low-frequency sub-band, mid-frequency sub-band and high-frequency sub-band. Differentiated fusion rules are used to fuse different frequency bands respectively, and then the fused sub-bands are reconstructed into an enhanced fused image. (3) Based on the dual-energy attenuation characteristics, calculate the R value (dual-energy attenuation ratio) for each pixel of the ore area image extracted after preprocessing. Based on the preset feature interval, calculate the proportion of pixels in the image whose R value falls in each feature interval. Combined with the calibration model of the mapping relationship between the pre-established R value and the element content inside the ore, output the grade estimation result. (4) Construct a sorting model. Using the YOLOv8x model framework, based on the fused image data and R-value analysis results, construct a dataset, train, optimize, and evaluate the model, and finally select the optimal sorting model. (5) Combine the output of the sorting model and the R-value analysis results to make a fusion decision and realize ore sorting.

[0005] Furthermore, the preprocessing described in step (1) includes: normalizing the acquired image, performing edge detection using the improved Canny algorithm, performing composite morphological processing and connected component analysis, accurately extracting the ore region, and eliminating background interference.

[0006] Furthermore, the improved Canny algorithm includes: High and low thresholds are automatically calculated based on image histogram analysis, with the high threshold set as follows: The low threshold is Where μ is the image mean, Standard deviation; Multi-scale Gaussian filtering, based on the estimated average ore diameter d avg Adaptive selection of multi-scale Gaussian filter kernel size; Gradient direction quantization quantizes the gradient direction into 8 main directions to improve computational efficiency.

[0007] Furthermore, the composite morphological processing includes first using a closing operation to connect the broken edges, then performing an opening operation to eliminate noise, and finally optimizing the contour integrity through adaptive iteration.

[0008] Further, in step (2), a two-dimensional discrete wavelet transform is used to decompose the ore region image into multiple frequency bands. Differentiated fusion rules are designed for different frequency bands: the fusion weight of the low-frequency sub-band is dynamically calculated by the fractal dimension and thickness compensation factor; the fusion mask of the mid-frequency sub-band is constructed based on local energy competition, and a brightness difference adjustment factor is introduced; the fusion weight of the high-frequency sub-band is calculated based on the gradient magnitude. Finally, the fused sub-bands are reconstructed into an enhanced fused image by inverse wavelet transform.

[0009] Furthermore, in step (3), the formula for calculating the pixel ratio is: P = N / N total Where N is the number of pixels whose R value falls within the feature interval, N total The total number of pixels in the ore region is denoted as . For ores with unknown content, the proportion of pixels whose R value falls within the feature interval in the ore region image is automatically calculated and substituted into a pre-established calibration model. This enables a rapid and non-destructive estimation of the content of lead-zinc ore and molybdenum ore. The calibration model is an empirical regression model, which is a mapping relationship between R value and ore grade established through known samples. It is used to convert the pixel proportion into grade estimation.

[0010] Further, in step (3), the formula for calculating the dual-energy attenuation ratio R is: ; Where V L V represents the pixel value of a low-energy image. H D represents the pixel value of a high-energy image. L For low-energy dark field values, D H For high-energy dark field values, G L For low-energy gain coefficients, G H This is the high-energy gain coefficient.

[0011] Further, in step (3), the method for determining the preset specific interval is as follows: based on the known composition of lead-zinc ore and molybdenum ore samples, a systematic calibration experiment is carried out. Through repeated imaging, statistical analysis and cross-validation, the preset characteristic interval of lead-zinc ore is determined to be 0.44-0.46, and the preset characteristic interval of molybdenum ore is 0.51-0.52.

[0012] Furthermore, when constructing the sorting model, the multiple images of molybdenum ore and lead-zinc ore after grade estimation in step (3) are divided into high and low content categories according to the set grade threshold. The data are labeled using tools to construct a high-quality dataset containing concentrate and tailings. The training set, validation set and test set are divided in an 8:1:1 ratio. Samples near the boundary value are cleaned to reduce noise interference.

[0013] Furthermore, the deep learning model adopts the YOLOv8x architecture with an input resolution of 640×640. It is optimized through loss functions CIoU Loss and Focal Loss, with 500 iterations. Finally, the training model with the best sorting effect is selected by calculating the model's accuracy and F1 score.

[0014] The present invention has at least the following beneficial effects: 1. The multi-band image fusion algorithm proposed in this invention, taking into account the different physical characteristics of molybdenum ore and lead-zinc ore, adopts different fusion parameters and discrimination criteria in the multi-band fusion process, which effectively enhances the effect of dual-energy image fusion of ore and overcomes the sorting problem of low-contrast ore. The sorting accuracy of zinc ore can reach 84%. Compared with the traditional fixed-weight fusion method, it has achieved significant improvements in core indicators such as precision, recall, mAP50 and mAP50-95.

[0015] 2. The R-value analysis method of the present invention calculates the pixel-level R-value of the dual-energy X-ray image of each ore, and counts the proportion of pixels in the image whose R-value falls within the feature interval. Finally, through a pre-established calibration model, this proportion is converted into an estimated ore content value, realizing rapid and non-destructive estimation of grade. At the same time, multiple molybdenum and lead-zinc ore images after R-value analysis and estimation are converted into high-quality labeled datasets, effectively solving the constraint of insufficient data samples on deep learning training models.

[0016] 3. The sorting method of the present invention effectively breaks through the bottlenecks of low efficiency and poor adaptability of traditional sorting technology. By optimizing and integrating the entire process of image acquisition, fusion, analysis and decision-making, it significantly improves the system's processing speed and recognition accuracy, and fully meets the high standards of real-time performance required in industrial settings. Attached Figure Description

[0017] Figure 1 These are comparison images of X-ray energy optimization, where Figure (a) is the energy image of 20kV-3.5mA, Figure (b) is the energy image of 160kV-1.2mA, and Figure (c) is the energy image of 160kV-3.5mA. Figure 2 This is a flowchart of dual-energy image preprocessing; Figure 3 Figure (a) is a schematic diagram of the image preprocessing results of an embodiment of the present invention, wherein Figure (b) is a cropped image of the lead-zinc ore region after morphological processing and background removal, and Figure (c) is a cropped image of the molybdenum ore region after morphological processing and background removal. Figure 4 These are schematic diagrams of dual-energy ore multi-band image fusion according to an embodiment of the present invention, wherein Figure (a) is a schematic diagram of the fusion effect of lead-zinc ore and Figure (b) is a schematic diagram of the fusion effect of molybdenum ore; Figure 5 This is a flowchart of R-value grade analysis; Figure 6 These are schematic diagrams illustrating the training effect of the YOLOv8 model in an embodiment of the present invention, wherein Figure (a) is a schematic diagram of the training effect of lead-zinc ore and Figure (b) is a schematic diagram of the training and beneficiation effect of molybdenum ore; Figure 7 These are schematic diagrams of the sorting results in an embodiment of the present invention, wherein Figure (a) is a schematic diagram of the sorting effect of lead-zinc ore and Figure (b) is a schematic diagram of the sorting effect of molybdenum ore; Figure 8 This is a performance comparison chart of the sorting results between the intelligent sorting method of this invention and the traditional fixed-weight fusion method. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0019] This invention provides an intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis. Through a series of precisely controlled image processing and data analysis steps, it achieves automatic identification and sorting of high-value ores such as lead-zinc ore and molybdenum ore. The intelligent mineral sorting method specifically includes: dual-energy X-ray image acquisition and preprocessing, multi-band image fusion, R-value calculation and grade analysis, sorting model construction, and fusion decision.

[0020] (1) Dual-energy X-ray image acquisition and preprocessing A dual-energy X-ray transmission imaging system was used to simultaneously acquire high-energy and low-energy images of the ore under optimized parameters. The acquired high-energy and low-energy images were then preprocessed to extract images of the ore region.

[0021] The image preprocessing stage includes image normalization, edge detection optimization, morphological processing, and ore region extraction. First, the 16-bit original image is linearly transformed to an 8-bit grayscale range, while retaining the original data for precise analysis. Then, an improved Canny algorithm is used to implement dynamic threshold calculation, multi-scale Gaussian filtering, and gradient direction quantization to enhance edge continuity. Next, closed-loop operations are used to connect broken edges, and open-loop operations are used to eliminate noise. An adaptive iterative strategy is then combined to optimize contour integrity. Finally, connected component analysis is used to accurately extract the ore region and effectively remove background interference.

[0022] This invention innovatively uses an improved Canny operator and composite morphological processing to accurately extract ore regions. Improved image preprocessing enhances the quality of the input image, thereby indirectly suppressing the interference of noise and matrix effects on subsequent R-value analysis, thus improving the effectiveness and accuracy of R-value estimation.

[0023] (2) Multi-band image fusion This invention uses an image fusion method based on multi-band decomposition. First, guided filtering is used to preprocess the high- and low-energy images of the extracted ore region to suppress noise while preserving the key edges and textures of the ore. Then, the image is decomposed into three layers through two-dimensional discrete wavelet transform to obtain the low-frequency subband (LL), the mid-frequency subband, and the high-frequency directional subband (LH, HL, HH).

[0024] Then, differentiated fusion rules are designed for different frequency bands: for low-frequency fusion, the fractal dimension (D) and thickness compensation factor (T) are combined to adaptively allocate weights and preserve the macroscopic structure of the ore; for mid-frequency fusion, a fusion mask (M) is constructed using local energy competition, and a brightness difference adjustment factor (W) is introduced to enhance the salient feature areas; for high-frequency fusion, adaptive weights are calculated based on gradient magnitude (G) to enhance fine edges while suppressing noise.

[0025] Finally, the fused subbands are reconstructed into an enhanced fused image by inverse wavelet transform, which significantly improves the distinction between the target minerals and the background.

[0026] (3) R-value calculation and grade analysis A method for calculating the R-value based on dual-energy attenuation characteristics is proposed for ore grade estimation. The R-value (dual-energy attenuation ratio) is calculated based on the pixel values ​​of dual-energy images, characterizing the differentiated attenuation of materials to X-rays. The proportion of pixels in the image whose R-values ​​fall within a preset feature range is statistically analyzed. Combined with a pre-established calibration model mapping the R-value to the elemental content within the ore, a rapid and non-destructive estimation of ore grade is achieved.

[0027] The preset feature range (e.g., molybdenum ore, lead-zinc ore) is determined through experimental analysis using ore samples with known composition. The calibration model is an empirical regression model, which is a mapping relationship between the R value and the ore grade established through known samples, used to convert pixel ratio into grade estimation.

[0028] (4) Construction of sorting model For multiple ore images estimated through R-value analysis, a high-quality dataset containing concentrate and tailings is constructed, and the training set, validation set, and test set are divided in an 8:1:1 ratio. The YOLOv8x model is adopted with an input resolution of 640×640. The model is optimized through loss functions CIoU Loss and Focal Loss, and the optimal model is selected to identify the fused images, outputting the spatial location, category, and confidence score of the ore, thus achieving accurate detection and classification of ore targets.

[0029] (5) Integrated decision-making The sorting model is responsible for target detection and preliminary classification, R-value analysis is used for grade estimation, and finally, the R-value analysis results and model output are combined to make a fusion decision based on category information and grade threshold, and control the sorting mechanism to achieve automatic ore sorting.

[0030] Implementation Cases This implementation example uses the above-mentioned intelligent mineral sorting method to intelligently sort molybdenum ore and lead-zinc ore, and compares its performance with that of the traditional fusion model to further verify the beneficial effects of the technical solution of the present invention.

[0031] 1. Dual-energy X-ray image acquisition and preprocessing X-ray source energy selection: In this case, considering the characteristics of lead-zinc ore, an adjustable X-ray generator was used to conduct a system experiment in the range of 80-160kV with a step size of 10kV. By evaluating image contrast, signal-to-noise ratio and boundary sharpness, the optimal imaging energy was determined. Finally, 160kV / 3.5mA was selected as the optimal parameter for subsequent dual-energy X-ray transmission and grade identification. Figure 1These are comparison images of optimized X-ray energy, where Figure (a) is the energy image of 20kV-3.5mA, Figure (b) is the energy image of 160kV-1.2mA, and Figure (c) is the energy image of 160kV-3.5mA.

[0032] Image preprocessing: such as Figure 2 The diagram shows the dual-energy image preprocessing process. This invention designs Canny edge detection + morphological operation to extract the ore region for preprocessing. By reading the original X-ray image, the image is divided into low-energy and high-energy parts. Edge extraction and image segmentation are performed, the ore region is extracted and the background is removed. Finally, the processed low-energy and high-energy images are saved respectively.

[0033] In practice, since the original image output by the X-ray detector used in this case is 16-bit grayscale (0-65535), while conventional edge detection algorithms are optimized for 8-bit images (0-255), the image needs to be linearly normalized to the 8-bit range first, while retaining the original 16-bit data for subsequent accurate analysis to ensure that no effective information is lost.

[0034] In the edge detection stage, an improved Canny algorithm is employed. Key improvements include: automatically calculating high and low thresholds based on image histogram analysis, with the high threshold set as follows: The low threshold is Where μ is the image mean, The standard deviation is given by the estimated average diameter d of the ore. avg The size of the multi-scale Gaussian filter kernel is adaptively selected, and the gradient direction is quantized into 8 main directions to improve computational efficiency.

[0035] To address the discontinuous nature of ore edges, a composite morphological processing flow was further designed: firstly, a closing operation was employed (…). Connect the fractured edges using elliptical structural elements with major and minor axis ratios set according to the ore shape. Expansion operations fill holes and gaps, while corrosion operations restore the main shape and preserve the breakpoint connections. Then, an opening operation is performed. To eliminate subtle noise, isolated points are removed by erosion and the main outline is restored by expansion. Finally, the number of iterations of morphological operations is dynamically adjusted based on the outline integrity analysis to optimize edge connectivity and noise suppression.

[0036] After the above processing, the ore region is extracted using connected component analysis, and the background is removed to obtain the following: Figure 3 The image shows high and low energy values ​​of the ore.

[0037] 2. Multi-band image fusion Gray-level histograms of dual-energy X-ray images of ore at high and low energy levels were acquired separately. Guided filtering was used for preprocessing to suppress noise and preserve edges. Then, two-dimensional discrete wavelet transform was used to decompose the images into low-frequency, mid-frequency and high-frequency sub-bands, thereby achieving effective fusion of molybdenum ore and lead-zinc ore.

[0038] In practice, differentiated fusion rules are applied to different frequency bands for fusion, mainly including three key steps: the fusion weights for the low-frequency sub-band are dynamically calculated based on the fractal dimension and thickness compensation factor; the fusion mask for the mid-frequency sub-band is generated based on local energy competition and brightness difference adjustment factors; and the weights for the high-frequency sub-band are assigned according to gradient amplitude to enhance edge details. Finally, the fused image is reconstructed through inverse wavelet transform, and the specific fusion effect is as follows: Figure 4 As shown.

[0039] 3. R-value calculation and grade analysis like Figure 5 As shown, the R-value grade analysis process is based on the physical mechanism of dual-energy X-ray imaging. By quantifying the differential attenuation characteristics of ore to high and low energy X-rays, a mapping relationship between the ore and its internal elemental content is established.

[0040] In practice, firstly, based on the Lambert-Beer law, the pixel grayscale values ​​of corresponding regions are extracted from the high-energy and low-energy images respectively. The attenuation ratio R is then calculated, defined as follows: ; In this formula, V L and V H D represents the transmission signals of the low-energy and high-energy channels, respectively. L and D H These are low-energy and high-energy dark field correction terms, used to eliminate detector noise, G L and G H These are the low-energy and high-energy gain correction coefficients, used to compensate for the response differences of different energy channels.

[0041] Dark field correction: The background value (noise) of the detector when there is no signal.

[0042] Gain correction: The ratio of the detector's response coefficient to different energy channels to its amplification coefficient.

[0043] This formula calculates the ratio after correcting high and low energy signals, thereby obtaining the R value that reflects the decay characteristics of the material. It effectively overcomes the influence of ore thickness variation on signal intensity and instead reflects the atomic number characteristics of the material, thus having a strong correlation with ore grade.

[0044] To establish the quantitative relationship between R-value and lead-zinc grade, this case study conducted a systematic calibration experiment based on 300 lead-zinc ore and molybdenum ore samples with known composition. Through repeated imaging, statistical analysis, and cross-validation, the characteristic R-value range applicable to lead-zinc ore was determined to be 0.44-0.46, and the characteristic R-value range applicable to molybdenum ore was determined to be 0.51-0.52.

[0045] Finally, for ores with unknown content, the content of lead-zinc ore and molybdenum ore can be quickly and non-destructively estimated by automatically calculating the proportion of pixels in the image whose R value falls within the above range and substituting them into a pre-established calibration model.

[0046] 4. Sorting Model Construction The sorting model is constructed based on dual-energy X-ray image data of molybdenum and lead-zinc mines, covering the entire process of data preparation, model training and performance evaluation.

[0047] In practice: For the 1335 molybdenum ore images and 1354 lead-zinc ore images whose ore content was estimated through R-value analysis, the ores were divided into high and low content categories according to the set grade threshold. The Labelimg tool was used for labeling, and the training set, validation set and test set were divided in an 8:1:1 ratio. Samples near the boundary value were cleaned to reduce noise interference.

[0048] The model was trained using the YOLOv8x architecture on a Windows system (Intel i9-12900H, NVIDIA RTX3060), with image resolution of 640×640, batch size of 32, and iterations of 500. The model's built-in enhancement module was utilized to avoid excessive manual enhancement. The training results for lead-zinc ore are as follows: Figure 6 As shown in (a), the training results for molybdenum ore are as follows: Figure 6 As shown in (b), the convergence of the loss function and the detection accuracy index (mAP) can verify that the model has good recognition ability and stability, and at the same time, it shows that the proposed image fusion and R-value analysis method can effectively improve the accuracy of ore sorting.

[0049] After training, the model was tested using 300 images of lead-zinc ore and molybdenum ore with known specific contents, respectively, and the results were as follows: Figure 7 The diagram shown illustrates the sorting results, in which... Figure 7 (a) shows the sorting effect of lead-zinc ore. Figure 7 (b) shows the sorting effect of molybdenum ore. Finally, by calculating the accuracy and F1 score of the model, the training model with better sorting effect was selected.

[0050] 5. Integrated decision-making By combining the output of the sorting model and the R-value analysis results, a fusion decision is made to achieve ore sorting.

[0051] Figure 8 The performance differences of the method of this invention and the traditional fixed-weight fusion method in terms of precision, recall, and mAP were compared. Quantitative analysis showed that the method of this invention exhibited significant advantages in core indicators such as precision, recall, mAP50, and mAP50-95. This further illustrates that multi-band adaptive fusion can effectively enhance image structure, texture, and detail information, improve the separability of ore features, and, combined with R-value analysis to introduce physical characteristic constraints, further improve the model's recognition accuracy and stability, thus verifying the effectiveness and superiority of this method in ore sorting.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart mineral sorting method based on dual-energy X-ray image fusion and R-value analysis, characterized in that, Includes the following steps: (1) High-energy and low-energy images of the ore were acquired simultaneously using a dual-energy X-ray transmission imaging system. The acquired images were preprocessed to extract images of the ore region. (2) After the extracted ore area image is processed by guided filtering, multi-band decomposition is performed to obtain low-frequency sub-band, mid-frequency sub-band and high-frequency sub-band. Differentiated fusion rules are used to fuse different frequency bands respectively, and then the fused sub-bands are reconstructed into an enhanced fused image. (3) Based on the dual-energy attenuation characteristics, calculate the R value (dual-energy attenuation ratio) for each pixel of the ore area image extracted after preprocessing. Based on the preset feature interval, calculate the proportion of pixels in the image whose R value falls in each feature interval. Combined with the calibration model of the mapping relationship between the pre-established R value and the element content inside the ore, output the grade estimation result. (4) Construct a sorting model. Using the YOLOv8x model framework, based on the fused image data and R-value analysis results, construct a dataset, train, optimize, and evaluate the model, and finally select the optimal sorting model. (5) Combine the output of the sorting model and the R-value analysis results to make a fusion decision and realize ore sorting.

2. The intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis according to claim 1, characterized in that, The preprocessing described in step (1) includes: normalizing the acquired image, performing edge detection using the improved Canny algorithm, performing composite morphological processing and connected component analysis, accurately extracting the ore region, and eliminating background interference.

3. The intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis according to claim 2, characterized in that, The improved Canny algorithm includes: High and low thresholds are automatically calculated based on image histogram analysis, with the high threshold set as follows: The low threshold is Where μ is the image mean, Standard deviation; Multi-scale Gaussian filtering, based on the estimated average ore diameter d avg Adaptive selection of multi-scale Gaussian filter kernel size; Gradient direction quantization quantizes the gradient direction into 8 main directions to improve computational efficiency.

4. The intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis according to claim 2, characterized in that, The composite morphological processing includes first using a closing operation to connect the broken edges, then performing an opening operation to eliminate noise, and finally optimizing the contour integrity through adaptive iteration.

5. The intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis according to claim 1, characterized in that, Step (2) uses two-dimensional discrete wavelet transform to decompose the ore region image into multiple frequency bands, and designs differentiated fusion rules for different frequency bands: the fusion weight of the low-frequency sub-band is dynamically calculated by fractal dimension and thickness compensation factor, the mid-frequency sub-band constructs a fusion mask based on local energy competition and introduces a brightness difference adjustment factor; the fusion weight of the high-frequency sub-band is calculated based on gradient magnitude. Finally, the fused sub-bands are reconstructed into an enhanced fused image by inverse wavelet transform.

6. The intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis according to claim 1, characterized in that, In step (3), the formula for calculating the pixel ratio is: P = N / N total Where N is the number of pixels whose R value falls within the feature interval, N total The total number of pixels in the ore region is denoted as . For ores with unknown content, the proportion of pixels whose R value falls within the feature interval in the ore region image is automatically calculated and substituted into a pre-established calibration model. This enables a rapid and non-destructive estimation of the content of lead-zinc ore and molybdenum ore. The calibration model is an empirical regression model, which is a mapping relationship between R value and ore grade established through known samples. It is used to convert the pixel proportion into grade estimation.

7. The intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis according to claim 1, characterized in that, In step (3), the formula for calculating the dual-energy attenuation ratio R is: ； Where V L V represents the pixel value of a low-energy image. H D represents the pixel value of a high-energy image. L For low-energy dark field values, D H For high-energy dark field values, G L For low-energy gain coefficients, G H This is the high-energy gain coefficient.

8. The intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis according to claim 1, characterized in that, In step (3), the method for determining the preset specific interval is as follows: based on the known composition of lead-zinc ore and molybdenum ore samples, a systematic calibration experiment is carried out. Through repeated imaging, statistical analysis and cross-validation, the preset characteristic interval of lead-zinc ore is determined to be 0.44-0.46, and the preset characteristic interval of molybdenum ore is 0.51-0.

52.

9. The intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis according to claim 1, characterized in that, When constructing the sorting model, the multiple images of molybdenum ore and lead-zinc ore after grade estimation in step (3) are divided into high and low content categories according to the set grade threshold. The data are labeled using tools to construct a high-quality dataset containing concentrate and tailings. The training set, validation set and test set are divided in an 8:1:1 ratio. Samples near the boundary value are cleaned to reduce noise interference.

10. The intelligent mineral sorting method based on dual-energy X-ray image fusion and R-value analysis according to claim 1, characterized in that, The deep learning model adopts the YOLOv8x architecture with an input resolution of 640×640. It is optimized through loss functions CIoULoss and Focal Loss, with 500 iterations. Finally, the training model with the best sorting effect is selected by calculating the model's accuracy and F1 score.

Images